Akademik Veri Yönetim Sistemi
13th INTERNATIONAL BILTEK CONGRESS ON CURRENT DEVELOPMENTS IN SCIENCE, TECHNOLOGY AND SOCIAL SCIENCES, Paris, Fransa, 18 - 21 Aralık 2025, ss.349-358, (Tam Metin Bildiri)
Predictor selection is a critical component of statistical downscaling, as the chosen atmospheric variables directly influence model performance and the quality of station-scale precipitation predictions. This study compares traditional correlation-based predictor selection with the All Possible Regression (APR) approach for monthly precipitation downscaling at the Bartın station. A total of 24 predictor variables from the ERA5 dataset were used, chosen based on their potential relationship with regional precipitation. In this regard, two selection methods were tested as a preliminary step for the GCM downscaling process. As the first approach, Pearson and Spearman correlations were used for predictor selection. The second approach involved evaluating all statistically feasible predictor subsets using APR, in which all possible predictor combinations were tested. For each subset, a MARS downscaling model was trained using 1979 2007 data and tested on 2007 2014. Results show that correlation-based selection is simple and easy to apply but often identifies redundant predictors and fails to capture variable interactions that real-world precipitation dynamics possess. APR-selected subsets demonstrated substantially higher predictive skill, yielding higher performance metrics for the test period and outperforming the correlation-based models. APR further distinguishes itself by performing predictor selection and pruning individually for each station, which is something correlation methods are not designed to do. These findings highlight APR as a superior predictor selection method for machine-learning-based downscaling and provide a strong foundation for future climate projection studies.